[llvm] [AArch64] Add support for 16/32/64-bit floating-point atomic read-modify-write ops (PR #125686)

[Tomas Matheson via llvm-commits]

================

----------------
tmatheson-arm wrote:

Please could you pre-commit the test changes in a separate PR, and include the sequences used when lsfe is not available.

https://github.com/llvm/llvm-project/pull/125686